Data Minimization

Every data element collected must be tied to an explicit, legitimate purpose declared before processing begins. This prevents function creep—where data collected for one task is silently repurposed for another.

• Article 5(1)(b) of GDPR mandates purpose limitation as a binding legal requirement
• Agents must validate that each API call or data fetch aligns with the declared processing purpose
• Purpose drift detection systems can flag when an agent requests data outside its operational mandate

Collect only what is adequate, relevant, and limited to the processing purpose. This directly counters model inversion risk by ensuring the model never sees extraneous sensitive attributes that could later be reconstructed.

• Strip quasi-identifiers (ZIP codes, birth dates, gender) when they aren't analytically necessary
• Implement field-level access controls that mask or redact attributes based on the agent's current task context
• Apply information gain analysis to verify each feature contributes measurable value before ingestion

Retain training data and inference logs only for the minimum duration required. Every stored record is a potential target for extraction attacks—reducing retention windows shrinks the blast radius.

• Implement time-to-live (TTL) policies on all agent memory stores and vector databases
• Automate secure deletion through cron-based data lifecycle managers that purge records past their retention period
• Differential privacy guarantees degrade over multiple queries; retention limits prevent long-term privacy budget exhaustion

Wherever possible, agents should operate on aggregated statistics rather than individual-level data. Aggregation prevents membership inference by removing the one-to-one mapping between outputs and specific training examples.

• Use k-anonymity thresholds: refuse queries that would return results from fewer than k individuals
• Replace raw data access with differentially private aggregate APIs that inject calibrated noise
• Agents performing analytics should request COUNT, SUM, or AVG operations rather than SELECT * queries

Before data enters any agent pipeline, validate and sanitize inputs to strip unnecessary precision and sensitive fields. This prevents accidental collection of high-resolution data that enables re-identification.

• Truncate timestamps to necessary granularity (e.g., month instead of millisecond)
• Round numerical values to reduce fingerprinting surface
• Reject or redact free-text fields containing personally identifiable information (PII) detected via named entity recognition

In federated learning, minimization means sharing only model updates, never raw data. Secure aggregation protocols further ensure the central server sees only the aggregated gradient, not individual contributions.

• Apply gradient clipping to bound the influence of any single data point
• Use secure aggregation to compute sums without exposing per-client updates
• Implement local differential privacy on-device before any data leaves the client boundary